[Openstack-operators] segmentation of management and data plane
Britt Houser (bhouser)
bhouser at cisco.com
Fri Sep 12 18:54:23 UTC 2014
Hey everyone - we've been doing some OpenStack deployments on different
datacenter architectures in our lab. Many of the architectures we work
with divide the management plane from the data plane by VRF. I.e. there
is no routing between a tenant VM traffic and the management interfaces of
the datacenter infrastructure. Applying this same model to OpenStack we
put the OpenStack services onto the management network. Tenants users are
given access to the management VRF so they can utilize the APIs. Tenant
VM traffic is kept in the data VRF. That works pretty well with a couple
So the first expect ion we noticed was the nova metadata service. Here,
the tenant VM needs access to an API on the management VRF. You can solve
this via config drive, or through neutron's metadata proxy. Not too bad.
The second exception we hit to this policy is Swift. Swift APIs actually
need to be accessed by Tenant VMs. Not a big deal, just put the swift
servers into the data management VRF. But if you are wanting to use Ceph,
and have Ceph be your storage for both Swift, Cinder, and Glance, now
you've got a problem b/c you need access to the cluster from both VRFs.
I'm just working through this stuff in my lab, so I'm hoping to get some
feedback from the real world. Has anyone setup their OpenStack cluster
with the management and data planes segmented by VRF? And did you have to
run into this or any other issue of needing traffic to cross between VRFs?
If so, how did you work around? Dual homed servers? Fusion router?
Something more elegant?
More information about the OpenStack-operators